The effect of spatial scale could not be ignored in identification results of forest types generated by multi-resolution images, and the influence of adding texture information from remote sensing data on the accuracy of forest trees species identification at different spatial resolutions has not been clearly addressed. To clarify this situation, we studied the Wangyedian forest farm in Northeast China, by using quasi-synchronous and geographical coordinate matched multi-resolution satellite observations (six spatial resolution levels: 1, 2, 4, 8, 16 and 30 m) which were supported with GF-1 PMS (pan and multi-spectra sensor), GF-2 PMS, GF-1 WFV (wide field view) and Landsat-8 OLI (operational land imager) and could investigate any possible correlations between spatial resolution and the recognition result, besides the influence of adding texture information. Five dominant tree species were classified and identified using Support Vector Machine (SVM) classifier. We also examined the identification results of the dominant forest trees species obtained by using the up-scaling algorithm. The results showed that overall classification accuracy of tree species was significantly influenced by the spatial resolution of images. The highest accuracy at the 4 m resolution, and the accuracy decreased to a minimum as the resolution reduced to 30 m. The addition of texture information increased classification accuracy using multispectral imagery with resolutions from 1 to 8 m, and the overall accuracy of dominant tree species identification created after adding texture information was 2.0%-3.6% higher than that from results of spectral information alone in the study area. However, the improvement of accuracy did not appear to hold for medium resolution imagery (16 and 30 m spatial resolution). In addition, there was a significant difference between the multi-scale classification results provided by up-scaled images and that obtained from native remote-sensing images for each spatial scale. These results indicated that the real satellite images should be used to ensure the accuracy of results when we examine multi-spatial-scale remote sensing observations or applications.
当前,不同空间分辨率卫星影像对森林类型识别结果中普遍存在的尺度效应,而且纹理参量对不同尺度下树种识别精度的影响仍缺乏广泛认知.本研究以中国东北旺业甸林场为研究区,采用观测时相同步、地理坐标匹配的GF-1 PMS、GF-2 PMS、GF-1 WFV,以及Landsat-8 OLI卫星传感器数据组成空间尺度观测序列(1、2、4、8、16、30 m),并结合支持向量机(SVM)模型,探讨了区域内5种优势树种遥感识别结果的尺度变化规律及其纹理特征参数的影响,同时检验了基于尺度上推转换影像的树种识别结果差异.结果表明: 影像空间分辨率对区域树种识别结果具有显著影响,其中,研究区森林树种识别的最佳影像分辨率为4 m,当分辨率降低至30 m时,树种识别结果最差.在1～8 m影像分辨率范围内,增加纹理信息能够显著提高不同优势树种的识别精度,使总分类精度提升了2.0%～3.6%,但纹理信息对16～30 m影像的识别结果没有显著影响.与真实尺度卫星影像相比,基于升尺度转换影像的树种识别结果及其尺度响应特征存在显著差异,表明在面向多个空间尺度的遥感观测和应用研究中,需要采用真实分辨率影像以确保结果的准确性.
Keywords: multispectral remote sensing; scale effect; spatial resolution; texture information; tree species classification.